Industrial Turbofan Predictive Maintenance & Remaining Useful Life (RUL) Forecaster

An end-to-end, production-grade machine learning system designed to predict the Remaining Useful Life (RUL) of commercial aircraft turbofan engines from continuous multi-sensor telemetry. Built on the NASA C-MAPSS dataset, this repository implements leak-free group cross-validation, asymmetric domain loss evaluation, game-theoretic SHAP explainability, and a deployable REST inference service.

---

📌 Executive Summary & Industrial Motivation

Unscheduled engine maintenance and in-flight shutdowns account for millions of dollars in carrier losses annually and introduce grave operational safety hazards. Traditional preventative maintenance follows rigid flight-hour schedules, either over-maintaining healthy turbines or missing accelerated failure propagation caused by harsh thermal cycles.

This project delivers an algorithmic condition-based prognostics engine that:

1. Detects subtle non-linear thermodynamic degradation across 21 sensor streams.
2. Formulates failure horizons via Piecewise Linear RUL Clipping ($RUL_{clip} = 125$), prioritizing model capacity where mechanical damage actively propagates.
3. Evaluates models through the NASA Asymmetric Scoring Function (PHM'08), penalizing hazardous late predictions exponentially more than conservative early ones.
4. Generates live additive feature attributions with SHAP, returning root-cause diagnostics alongside every RUL inference.

---

🏛️ End-to-End Pipeline Architecture

flowchart TD
    subgraph Data Ingestion
        A[Raw NASA C-MAPSS Telemetry .txt] --> B[src.data_loader: Canonical Schema & Validation]
    end

    subgraph Preprocessing & Feature Engineering
        B --> C[src.preprocessor: Piecewise Linear RUL Clipping at 125]
        C --> D[Sensor Variance Screening: Drop s1, s5, s6, s10, s16, s18, s19]
        D --> E[src.feature_engineer: Rolling Window Stats w=10 per Engine]
        D --> F[Cumulative Absolute Variation Wear Proxy]
        D --> G[Normalized Cycle Position]
        E & F & G --> H[Feature Matrix: 85 Engineered Features]
        H --> I[MinMaxScaler: Fit Strictly on Train]
    end

    subgraph Modeling & Cross-Validation
        I --> J[5-Fold GroupKFold CV on engine_id]
        J --> K[Ridge Regression]
        J --> L[Random Forest]
        J --> M[XGBoost]
        J --> N[LightGBM + Optuna Tuning]
        K & L & M & N --> O[Evaluator: RMSE, MAE, NASA PHM08 Penalty]
    end

    subgraph Deployment & Explainability
        N --> P[Champion Model Serialized to models/]
        P --> Q[SHAP TreeExplainer: Beeswarm & Local Waterfall]
        P & I & Q --> R[Antigravity REST API: POST /predict]
        R --> S[Real-Time RUL + Alerts + Top 3 SHAP Drivers]
    end

Loading

---

🔬 Dataset Overview (NASA C-MAPSS)

The Commercial Modular Aero-Propulsion System Simulation benchmark encompasses four subsets simulating complex failure propagation:

Subset	Train Engines	Test Engines	Operating Regimes	Fault Modes	Primary Challenge
FD001	100	100	1 (Sea Level)	1 (HPC Degradation)	Baseline single-mode prognostics
FD002	260	259	6 (Flight Envelopes)	1 (HPC Degradation)	Multi-regime normalization & clustering
FD003	100	100	1 (Sea Level)	2 (HPC + Fan)	Complex multi-fault interaction
FD004	248	249	6 (Flight Envelopes)	2 (HPC + Fan)	Combined regime shifts & multi-component wear

---

💡 Key Methodological Differentiators

1. Leak-Free GroupKFold Cross-Validation

Standard random train_test_split or regular KFold causes catastrophic data leakage in time-series prognostics, because cycle $t$ and cycle $t+1$ of the same engine share identical baseline geometry. We strictly enforce 5-Fold GroupKFold grouped by engine_id, guaranteeing zero telemetry bleeding between folds.

2. Piecewise Linear RUL Target Formulation ($RUL_{clip} = 125$)

Early in an engine's lifecycle, mechanical wear is physically undetectable. Attempting to differentiate between 280 and 320 cycles remaining penalizes models for learning arbitrary sensor noise. Capping $RUL \le 125$ focuses gradient optimization exclusively onto the degradation inflection zone.

3. NASA Asymmetric Scoring Function

Standard RMSE treats overestimating and underestimating RUL symmetrically. In aviation, an overestimation (late prediction) is catastrophic (predicting 20 cycles when only 5 remain leads to in-flight engine failure), whereas an underestimation (early prediction) merely prompts premature inspection.

$$d_i = \hat{y}_i - y_i$$

$$S = \sum_{i=1}^{N} s_i, \quad s_i = \begin{cases} \exp\left(-\frac{d_i}{13}\right) - 1, & d_i < 0 \text{ (Early)} \ \exp\left(\frac{d_i}{10}\right) - 1, & d_i \ge 0 \text{ (Late)} \end{cases}$$

4. Explainability as a First-Class API Service

Predictions are not delivered as isolated scalars. Every API response computes additive Shapley values, categorizing whether physical sensor shifts (e.g. rising HPT outlet temperatures or fluctuating HPC pressure) are accelerating or prolonging turbine life.

---

📊 Benchmarking & Model Evaluation

Rigorous 5-fold cross-validation on FD001 (20,631 cycles) across linear and gradient-boosted architectures:

Architecture	CV RMSE (Cycles)	CV MAE (Cycles)	CV NASA Penalty	Training Time
LightGBM (Optuna Tuned)	16.16	11.27	111,189.5	2.22s
LightGBM	16.57	11.21	124,648.7	2.60s
XGBoost	16.63	11.26	126,522.5	9.13s
Random Forest	16.80	11.31	127,866.5	119.4s
Ridge Regression	22.22	17.67	248,866.3	0.25s

(Results from 5-fold leak-free GroupKFold CV on FD001 training set)

Unseen Test Set Evaluation (100 Engines at Terminal Observed Cycle)

Evaluated against ground truth RUL_FD001.txt:

• Test RMSE: 18.80 cycles
• Test MAE: 13.62 cycles
• Test MAPE: 25.72%
• Test NASA Score: 694.7

---

🔍 Explainability & Diagnostic Insights

Global Importance (SHAP Beeswarm)

Analysis of the top feature attributions reveals:

• s12_mean_10 (Ratio of Fuel Flow to Ps30): Dominant leading indicator. Elevated fuel consumption to maintain core pressure signals severe compressor fouling.
• s11_mean_10 & s11_std_10 (HPC Static Pressure): Increased pressure variance flags aerodynamic stall flutter prior to mechanical fatigue.
• cycle_norm & s2_cum_change: Monotonic wear accumulators anchor base degradation trajectory.

Local Root-Cause Diagnosis (Single Engine Waterfall)

For an engine entering critical alert status:

---

🚀 Antigravity REST API Service

Starting the Live Server

python api/app.py

Health Check

curl -X GET http://localhost:5000/health

{
  "status": "ok",
  "model_loaded": true,
  "framework": "Antigravity REST"
}

Predictive Inference with Live SHAP Attributions

curl -X POST http://localhost:5000/predict \
  -H "Content-Type: application/json" \
  -d '{
    "engine_id": 42,
    "cycles": [
      {
        "cycle": 150,
        "op_setting_1": -0.0007,
        "op_setting_2": -0.0004,
        "op_setting_3": 100.0,
        "s2": 641.82, "s3": 1589.70, "s4": 1400.60, "s7": 554.36,
        "s8": 2388.06, "s9": 9046.19, "s11": 47.47, "s12": 521.66,
        "s13": 2388.02, "s14": 8138.62, "s15": 8.4195, "s17": 392,
        "s20": 39.06, "s21": 23.4190
      }
    ]
  }'

Response Payload:

{
  "engine_id": 42,
  "predicted_rul": 37.2,
  "alert_level": "WARNING",
  "confidence_note": "Within degradation zone (RUL <= 50)",
  "top_shap_contributors": [
    {
      "feature": "s12_mean_10",
      "shap_value": -18.42,
      "direction": "accelerates_failure"
    },
    {
      "feature": "s7_std_10",
      "shap_value": -11.23,
      "direction": "accelerates_failure"
    },
    {
      "feature": "cycle_norm",
      "shap_value": -8.91,
      "direction": "accelerates_failure"
    }
  ]
}

Industrial Alert Thresholds:

Remaining Useful Life	Alert Level	Operational Protocol
$\text{RUL} > 100$ cycles	HEALTHY	Routine monitoring; nominal flight scheduling
$50 < \text{RUL} \le 100$	WATCH	Flag engine for non-destructive inspection at next hub
$20 < \text{RUL} \le 50$	WARNING	Restrict flight envelopes; schedule module overhaul
$\text{RUL} \le 20$	CRITICAL	Immediate grounding; engine swap required

---

🧪 Automated Testing

Comprehensive test suite covering preprocessor logic, rolling window isolation, metric asymmetry, and REST endpoints:

python -m pytest tests/ -v

tests/test_api.py::test_health_endpoint PASSED                           [  7%]
tests/test_api.py::test_predict_empty_payload PASSED                     [ 14%]
tests/test_api.py::test_predict_empty_cycles PASSED                      [ 21%]
tests/test_api.py::test_predict_successful_mock PASSED                   [ 28%]
tests/test_feature_engineer.py::test_add_rolling_features_isolation PASSED [ 35%]
tests/test_feature_engineer.py::test_add_cumulative_wear_features PASSED [ 42%]
tests/test_feature_engineer.py::test_add_normalized_cycle PASSED         [ 50%]
tests/test_feature_engineer.py::test_operational_clusters PASSED         [ 57%]
tests/test_preprocessor.py::test_add_rul_labels_unclipped PASSED         [ 64%]
tests/test_preprocessor.py::test_add_rul_labels_clipped PASSED           [ 71%]
tests/test_preprocessor.py::test_add_test_rul_labels PASSED              [ 78%]
tests/test_preprocessor.py::test_identify_low_variance_features PASSED   [ 85%]
tests/test_preprocessor.py::test_drop_features PASSED                    [ 92%]
tests/test_preprocessor.py::test_fit_and_transform_scaler PASSED         [100%]

---

📂 Repository Structure

.
├── data/
│   ├── raw/                       # NASA C-MAPSS raw .txt files (FD001 - FD004)
│   └── processed/                 # Engineered feature matrices (.parquet)
├── notebooks/
│   ├── 01_EDA.ipynb               # Sensor variance, distributions, correlation
│   ├── 02_Feature_Engineering.ipynb # Rolling windows, wear proxies, clipping
│   ├── 03_Model_Benchmarking.ipynb  # GroupKFold CV, Optuna tuning, leaderboards
│   └── 04_SHAP_Explainability.ipynb # Beeswarm, waterfalls, live attribution tests
├── src/
│   ├── __init__.py
│   ├── data_loader.py             # Schema-enforced ingestion & RUL loader
│   ├── preprocessor.py            # Labeling, clipping, variance filter, scaling
│   ├── feature_engineer.py        # Temporal stats, wear proxy, KMeans clustering
│   ├── trainer.py                 # GroupKFold CV, model instantiators, Optuna
│   ├── evaluator.py               # RMSE, MAE, NASA Asymmetric Scoring Function
│   └── explainer.py               # SHAP TreeExplainer wrapper & figure exports
├── api/
│   ├── __init__.py
│   ├── app.py                     # Antigravity REST API application
│   ├── schemas.py                 # Request/response schemas & alert logic
│   ├── predict.py                 # Production inference engine pipeline
│   └── requirements_api.txt
├── models/
│   ├── best_model.pkl             # Serialized champion model (LightGBM)
│   ├── scaler.pkl                 # Fitted MinMaxScaler on training features
│   └── feature_columns.pkl        # Canonical feature ordering
├── reports/
│   ├── figures/                   # High-res publication plots & SHAP visuals
│   └── model_comparison.csv       # Benchmark results across all models
├── tests/
│   ├── __init__.py
│   ├── test_api.py                # REST endpoint integration tests
│   ├── test_feature_engineer.py   # Leakage prevention & isolation tests
│   └── test_preprocessor.py       # RUL clipping and scaling tests
├── requirements.txt
├── pytest.ini
└── README.md

---

📚 References & Academic Citation

• Saxena, A., Goebel, K., Simon, D., & Eklund, N. (2008). Damage Propagation Modeling for Aircraft Engine Run-to-Failure Simulation. In Proceedings of the 1st International Conference on Prognostics and Health Management (PHM08), Denver CO.
• Saxena, A., & Goebel, K. (2008). Turbofan Engine Degradation Simulation Data Set. NASA Ames Prognostics Data Repository, NASA Ames Research Center, Moffett Field, CA.
• Lundberg, S. M., & Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems (NeurIPS 2017).